feat(hint): H2 Phase B — per-form keying + the three exemplars (ADR-0053)

The first exemplar (`add 1:n relationship`) showed per-node keying is
too coarse for multi-form commands, so revise the mechanism to per-form.

- CommandNode `hint_id: Option<&str>` -> `hint_ids: &[&str]` (mirrors
  usage_ids); hint_key_for_input_in_mode reuses a factored-out
  pick_form_key (shared digit/m:n/suffix form disambiguation with
  usage_key_for_input_in_mode)
- wire INSERT + ADD (all four forms) with hint_ids
- author the three approved exemplars: hint.cmd.insert,
  hint.cmd.add_relationship, hint.err.foreign_key.child_side
  (what/example/concept) + keys.rs registration
- revise ADR-0053 D3 to per-form; record clause-concept hints as a
  deferred extension (issue #37); update README + plan
- +5 tests; 2488 pass / 1 ignored, clippy clean

docs/adr/0053-contextual-hint-command-and-keybinding.md

@@ -2,14 +2,18 @@
 
 ## Status
 
-Accepted — implementation pending. Revised after a `/runda` review
+Accepted — implementation in progress. Revised after a `/runda` review
 (2026-06-14): corrected the verbosity-default fact; re-keyed tier-3
-content on a new `hint_id` (not `help_id`) so every command form — simple
-and advanced-SQL — gets distinct, mode-correct content; split the
-pre-submit-diagnostic and runtime-error paths; added a comprehensiveness
-coverage test. The parallel question of whether the in-app `help` command
-should likewise distinguish advanced-SQL forms is tracked **separately**
-as Gitea issue #36 (it touches shipped, ADR-backed `help` behaviour).
+content off `help_id`; split the pre-submit-diagnostic and runtime-error
+paths; added a comprehensiveness coverage test. Revised again during
+Phase B implementation (2026-06-15): the first exemplar showed per-*node*
+keying is too coarse for multi-form commands (`add`/`drop`/`show`/
+`create`), so D3 now keys tier-3 content **per form** via a
+`hint_ids: &[&str]` array mirroring `usage_ids` — and **clause-concept
+hints** are recorded as a deferred extension (separate tracking issue).
+The parallel question of whether the in-app `help` command should
+likewise distinguish advanced-SQL forms is tracked **separately** as
+Gitea issue #36 (it touches shipped, ADR-backed `help` behaviour).
 
 Decided in conversation 2026-06-14. Closes the last open piece of **A1**
 (the canonical app-command set, ADR-0003): every app command is
@@ -132,22 +136,42 @@ top-level namespace (where tier-2 ambient strings already live), in two
 new sub-namespaces:
 
 - **`hint.cmd.<hint_id>`** — one per command **form**, keyed by a **new
-  `hint_id: Option<&'static str>`** field added to `CommandNode`
-  (`src/dsl/grammar/mod.rs:512`, parallel to the existing `help_id` /
-  `usage_ids`). The F1 live-input path resolves the current input to its
-  command node and looks up `hint.cmd.<node.hint_id>`.
+  `hint_ids: &'static [&'static str]`** field on `CommandNode`
+  (`src/dsl/grammar/mod.rs:512`), **mirroring the existing `usage_ids`**.
+  The F1 live-input path resolves the current input to its form's hint key
+  via `hint_key_for_input_in_mode`, which reuses the same form-word
+  disambiguation as `usage_key_for_input_in_mode`.
 
-  **Why a new field, not `help_id`:** `help_id` is **not** 1:1 with
-  command forms. The 7 advanced-mode SQL nodes (`SELECT`, `WITH`,
-  `SQL_INSERT/UPDATE/DELETE`, `EXPLAIN_SQL`) carry `help_id: None` *purely
-  to dedup the `help` command's printed list* (they share an entry word
-  with a simple sibling — see `grammar/mod.rs:915-918`), not because they
-  lack distinct content. Their SQL syntax differs from the simple-DSL
-  sibling's, so they **must get their own tier-3 block**. A dedicated
-  `hint_id` gives every one of the ~37 REGISTRY nodes — simple and
-  advanced-SQL alike — its own key and its own mode-correct example, with
-  no sharing or deferral. (The analogous gap in the `help` command is out
-  of scope here — issue #36.)
+  **Why an array mirroring `usage_ids`, not a per-node `hint_id`**
+  *(`/runda`/implementation revision, 2026-06-15)*: a single per-node key
+  is too coarse. Several entry words are **one node spanning many forms** —
+  `add` (column/relationship/index/constraint), `drop` (table/column/
+  relationship/index), `show` (data/table/tables/relationships/indexes),
+  `create` (table/index). A live-input hint for `add 1:n relationship` is
+  only useful if it is *specific to relationships*, so the content must be
+  **per form**, not per node. The project already solved exactly this for
+  usage templates (`usage_ids` is a per-form array, disambiguated by the
+  form word), so `hint_ids` mirrors it. Single-form nodes carry one entry;
+  multi-form nodes carry one per form. This also covers the advanced-SQL
+  forms whose `usage_ids` are empty (`SQL_INSERT/UPDATE/DELETE`,
+  `EXPLAIN_SQL`) — they get their own `hint_ids` directly, independent of
+  usage, with mode-correct SQL examples. (The `help`-list collapse of
+  advanced-SQL forms is a separate gap — issue #36.)
+
+  **Deferred extension — clause-concept hints** (issue #37): per-form is
+  the right granularity for tier-3 *teaching* (position-awareness within a
+  form is owned by tier-2 ambient + the live `Next:` line, D4). But some
+  **concepts live inside a clause**, not a form — `… on delete ⟨cascade|
+  set null|restrict⟩` (referential actions), the `create table` constraint
+  slots (`primary`/`unique`/`check`/`foreign`), `with pk`, `1:n`/`m:n`
+  cardinality. A learner parked in such a clause may want teaching deeper
+  than tier-2's candidate list but narrower than the whole-form block. v1
+  does **not** build this (it would multiply content for points whose value
+  we can't yet measure, and we don't expect to accumulate usage statistics
+  to drive it empirically — it will be tackled as a deliberate follow-up
+  job). The keying does not lock it out: a later `hint.concept.<topic>`
+  namespace can be surfaced when the cursor sits in a recognized clause,
+  layered on top of the per-form block.
 - **`hint.err.<class>`** — one per error/diagnostic class, keyed by the
   friendly error/diagnostic key (e.g. `hint.err.foreign_key.child_side`,
   `hint.err.type_mismatch`, `hint.err.insert_arity_mismatch`). Used by
@@ -309,10 +333,12 @@ Hint — add relationship
   the ambient one-liner, and the verbose error hint — without cluttering
   those terse defaults.
 - **One new keybinding (F1)** joins the keymap and the ADR-0051 strip.
-- **A new `hint_id` field on `CommandNode`** (parallel to `help_id`), one
-  new field of `App` state (`last_error_hint_key`), and one new renderer
-  family (`note_hint*`); the `AppCommand` enum gains `Hint`, the grammar a
-  `HINT` node, the REGISTRY one entry.
+- **A new `hint_ids: &[&str]` field on `CommandNode`** (mirroring
+  `usage_ids`) + a `hint_key_for_input_in_mode` lookup (reusing the
+  `usage_key_for_input_in_mode` form-disambiguation), one new field of
+  `App` state (`last_error_hint_key`), and one new renderer family
+  (`note_hint*`); the `AppCommand` enum gains `Hint`, the grammar a `HINT`
+  node, the REGISTRY one entry.
 - **A large, durable content corpus** (~37 command blocks + ~42 error/
   diagnostic blocks ≈ 80) enters the catalogue under `hint.cmd.*` /
   `hint.err.*`, validated by `keys.rs`. This is ongoing surface area: new
@@ -349,6 +375,12 @@ Hint — add relationship
   created a table — here's what an index would add") — OOS (deferred): a
   plausible future tier-3 use, but v1 scopes the command path to errors
   and the F1 path to in-progress input.
+- **Clause-concept hints** (`… on delete ⟨action⟩`, constraint slots,
+  `with pk`, cardinality) — OOS (deferred, issue #37): a
+  `hint.concept.<topic>` layer surfaced when the cursor sits in a
+  recognized clause, deeper than tier-2's candidate list but narrower than
+  the per-form block. Per-form keying (D3) does not lock it out. To be
+  tackled as a deliberate follow-up job, not gated on usage statistics.
 
 ## Content inventory (implementation tracking)
 

docs/adr/README.md

@@ -58,4 +58,4 @@ This directory contains the project's ADRs, recorded per
 - [ADR-0050 — Incidental-DDL confirmations omit relationship info (structure-only)](0050-incidental-ddl-confirmations-omit-relationships.md) — **Accepted + implemented 2026-06-12 (issue #28)**, closes Gitea **#28**. **Supersedes** the incidental-DDL clause of **ADR-0044 §1** and the relationship-block half of **ADR-0016 §5**. Incidental-DDL confirmation echoes (`create table`, `add`/`drop`/`rename`/`change column`, `add`/`drop index`) now render **structure only** — header + column box + `Indexes:` + constraints — with **no `References:` / `Referenced by:` block** (neither prose nor diagram), even when the table carries relationships the user did not touch. Rationale (owner): a confirmation echo reports the change just made, not untouched relationships; ADR-0044's terse prose was the lesser of "prose vs diagram", but the right answer for these surfaces is **neither**. **Relationship-subject surfaces are unchanged** — `show table`, `add`/`drop relationship`, `show relationship` still render ADR-0044 diagrams; relationships appear only when the user asks for (`show table`) or acts on (`add`/`drop relationship`) one, and are one `show table <T>` away — **no information lost**. Forks both user-chosen: **scope = all incidental DDL** (not just `add column` — the rationale is uniform, the mental model clean, and it's the simpler edit) and **delete the prose renderer** (not retain it dormant — no dead code). **Mechanism:** the `handle_dsl_success` `matches!` routing is unchanged (relationship-subject → diagrams; else → `render_structure`); the change is one line inside `render_structure` (`output_render.rs` — drop the relationship-block call) since all its callers are incidental DDL, plus deletion of the orphaned `relationship_prose_lines` + `cols_disp` helpers. The prose format survives in ADR-0016 §5 + git history for a future OOS-7 always-prose setting. **Tests:** the prose-presence unit test + its snapshot removed; a new unit test asserts `render_structure` on a description carrying **both** inbound and outbound relationships emits the box but no prose; the misnamed `add_relationship_flow_shows_inbound_section_on_parent` integration test (which sent an `AddColumn`) inverted + renamed to assert the add-column echo omits the prose; the diagram tests (`show table`, `add relationship`) unaffected. **2458 pass / 0 fail / 0 skip (1 ignored), clippy clean**. `requirements.md` unaffected (ADR-tracked refinement of a decided area, like ADR-0044 itself)
 - [ADR-0051 — Bottom keybinding strip: context- and state-aware](0051-context-state-aware-keybinding-strip.md) — **Accepted + implemented 2026-06-13 (issue #27)**, closes Gitea **#27**. Repurposes the bottom status line into a **keystrokes-only, state-selected** strip (builds on ADR-0046 nav focus, ADR-0003 modes, ADR-0049 the #29 readline keys it now advertises, ADR-0022 the completion memo). A pure `status_bar_bindings(app) -> Vec<(key,label)>` chooses the strip by **priority, first match wins**: (1) **sidebar focus** → `Ctrl-O next pane · ↑↓/PgUp/PgDn scroll · Esc input`; (2) **completion memo live** (`last_completion`) → `Tab/Shift-Tab cycle · Esc cancel · Enter run`; (3) **history navigation** (new `App::is_browsing_history()` exposing the private `history_cursor`) → `↑↓ browse · Esc clear · Enter run`; (4) **editing** (input non-empty) → `Esc clear · Ctrl-A/E home/end · Ctrl-W del word · Enter run` (surfaces the #29 keys, closing ADR-0049's deferred advertisement); (5) **default** (empty) → `Ctrl-O sidebar · Tab complete · ↑ history · Enter run`. Priority is correct because Up clears the completion memo and Tab cancels history nav, so states 2/3 never co-occur, and the five are exhaustive for Input focus. **Typed-command words leave the strip** (`mode advanced`/`mode simple` switch, `:` one-shot) and **mode discovery moves to the empty-input hint** (`resolve_hint_lines`), **simple mode only**: `\`mode advanced\` for SQL` (the verb "type" omitted — the prompt implies it; advanced mode shows **no** pointer per a post-trial user decision — a switcher knows how they got there and `help` covers the way back). The one-shot's old `Backspace cancel one-shot` label is subsumed by the editing state (behaviour intact). Forks all user-chosen: **editing state shows the #29 keys** (vs unadvertised); **`Ctrl-C quit` omitted** from the strip (vs always shown); **no width-drop machinery** — the longest strip (~65 cols) fits all supported widths, so a **width-budget unit test** keeps it lean by construction instead (the user's own observation). Catalog: 12 new `shortcut.*` labels + the `panel.hint_mode_advanced` string added to `en-US.yaml`+`keys.rs` (validator-checked 1:1), 5 now-dead strip strings removed. **Modal-aware strip is OOS** (pre-existing: a modal owns the keyboard and carries its own hints; the strip under it is unchanged-in-kind, not worsened). Tests: 9 Tier-1 unit (per-state key sets — completion/history driven through real key events; width budget; mode-pointer presence/absence), 1 Tier-3 rewritten (`status_bar_is_keystroke_only_and_state_aware`), 15 full-panel snapshots re-accepted (reviewed — strip/hint only). **2467 pass / 0 fail / 0 skip (1 ignored), clippy clean.** OOS: modal-aware strip; a full-key cheatsheet overlay; Ctrl-K/U advertisement (editing strip shows the highest-value subset within the width budget)
 - [ADR-0052 — Mode-tagged history for cross-mode recall](0052-mode-tagged-history-cross-mode-recall.md) — **Accepted + implemented 2026-06-13 (issue #30)**, closes Gitea **#30** — the feature (advanced history reusable in simple mode) **and** the bug in its comment (the `:` one-shot prefix lost across sessions). **Amends ADR-0034** (status field gains a `:adv` tag; **journaling moves from the worker to the dispatch layer**), **ADR-0015 §5/§6** (history.log leaves the worker transaction — `commit-db-last` now scopes yaml/csv/db only), and **ADR-0040** (a success-path journal-write failure is best-effort, not fatal); references ADR-0003. **Root cause:** history carried no mode, and the in-memory ring stored the raw `:select 1` while the worker journalled the *stripped* `select 1`, so the `:` was lost on disk. **Fix:** record the submission mode per entry as a **`:adv` suffix on the status token** (`ok`/`ok:adv`/`err`/`err:adv`) — `source` stays last + canonical so replay is unaffected; the in-memory ring (still `Vec<String>`) stores advanced entries in their `: `-prefixed simple-mode runnable form (a leading `:` unambiguously marks advanced since simple DSL never starts with `:`); recall **strips the `:` in advanced mode** (runs as bare SQL) and keeps it in simple mode (runs via the one-shot escape); hydration reconstructs the `: `-prefix from the tag, so cross-session = in-session. **The architectural turn (user's call):** the first draft kept journaling in the worker + threaded the mode down (~30-site plumbing); on review the user asked why the journal is written deep in the worker when the *failure* path already journals at the top of the chain — it shouldn't (history.log is a journal, not state). So **success journaling moved up** to `spawn_dsl_dispatch` / `run_replay` / the app-command sites (next to the failure path), the worker's `finalize_persistence` now writes only yaml/csv, and the journal write became **best-effort** (the command is already committed — consistent with the failure path; a rare disk-full leaves a committed command unjournalled, state intact). **App commands** journal simple (dispatched outside the spawn) and `submit` excludes them from the ring's advanced flag, so `undo`/`mode advanced` recall bare. Forks user-chosen: status-tag format (vs 4th field / `:`-in-source); unified scope; **dispatch-layer best-effort journaling** (vs worker-coupled-fatal). Two `/runda` passes (the second drove the relocation + app-command exclusion). Tests: the 15 worker-level journaling tests retired (worker no longer journals — yaml/csv/operation checks kept), re-covered at the new layer (history.rs status-tag + `:`-reconstruct; app.rs recall matrix; the #30 cross-session regression in `iteration6`; replay tests cover `run_replay` journaling). **2471 pass / 0 fail / 0 skip (1 ignored), clippy clean.** replay re-journaling mode-fidelity (a replayed advanced line re-journals simple — not a regression). **Follow-up done 2026-06-14:** the vestigial worker `source` plumbing was fully unwound (compiler-guided, no behaviour change) — `_source` removed from `finalize_persistence`/`do_rebuild_from_text`, the three `*_request` wrappers inlined+deleted, the dead `source` param dropped from the ~30 forwarding worker handlers, and the `source` field removed from the `DescribeTable`/`QueryData`/`RunSelect` requests + their `DatabaseHandle` methods (~164 mostly-test call sites); the only worker `source` left is the snapshot/undo label (see ADR-0052 *Consequences*)
-- [ADR-0053 — Contextual `hint` command and keybinding](0053-contextual-hint-command-and-keybinding.md) — **Accepted, implementation pending (2026-06-14)**. Settles the `hint` slot ADR-0003 left "ADR pending"; closes the last open piece of **A1** and tracks requirements **H2**. **Two surfaces:** an **F1 keybinding** that renders a deep hint for the *live* partial input without submitting (the primary path — a submitted `hint` command can't see the buffer it would help with, since Enter empties it), and a submitted **`hint` command** that expands on the *most recent error*. **No topic argument** (contextual only — `help <topic>` already owns explicit reference). Introduces a **tier-3 teaching layer**, deeper than the existing tier-1 (colour / error headline) and tier-2 (ambient one-liner; and the error `hint:`, which is shown **by default** since `Verbosity::Verbose` is the default — `messages short` is the opt-*out*); without it `hint` would just duplicate what's already on screen. Tier-3 content lives in the catalogue under `hint.cmd.<hint_id>` (per command form) and `hint.err.<class>` (per error/diagnostic class), each a structured `what`/`example`/`concept` block rendered via a new `note_hint*` family with `OutputStyleClass::Hint`. **Keyed on a new `hint_id` field on `CommandNode`, not `help_id`** (`/runda` correction): `help_id` is not 1:1 with command forms — the 7 advanced-mode SQL nodes carry `help_id: None` purely to dedup the `help` *list*, so they'd be unkeyable and would wrongly share their simple sibling's content despite different syntax; a dedicated `hint_id` gives every one of the ~37 REGISTRY nodes its own mode-correct block (the parallel `help`-side gap is tracked as issue **#36**). Two error routes share `hint.err.*`: pre-submit `diagnostic.*` read live from the walker (F1 path), runtime `translate_error` classes via stored `last_error_hint_key` (`hint` command / empty-F1). Adds `AppCommand::Hint`, a `HINT` grammar node + REGISTRY entry, the `hint_id` field, and `last_error_hint_key`; F1 is a read-only overlay (buffer + completion memo untouched). **Content is the bulk of the work** (the mechanism is ~a day): **comprehensive for v1** — ~37 command forms + 9 runtime error classes + ~33 `diagnostic.*` classes ≈ 80 teaching blocks — authored **exemplars-first** (voice approved in this ADR's `/runda` review, then mass-authored in batches), enforced by a **comprehensiveness coverage test** (every node/error class has a key), with graceful fall-back to tier-2 if a key is ever missing. Forks user-chosen: two-surface model; **F1** (vs `?` / a chord); no-arg; comprehensive scope; exemplars-first. OOS: per-topic `hint <topic>` (rejected — overlaps `help`); always-on tier-3 (rejected — keeps ambient terse); non-`en-US` locales + success-command teaching (deferred); the `help`-side advanced-SQL gap (issue #36)
+- [ADR-0053 — Contextual `hint` command and keybinding](0053-contextual-hint-command-and-keybinding.md) — **Accepted, implementation in progress (2026-06-14; Phase A done, Phase B underway)**. Settles the `hint` slot ADR-0003 left "ADR pending"; closes the last open piece of **A1** and tracks requirements **H2**. **Two surfaces:** an **F1 keybinding** that renders a deep hint for the *live* partial input without submitting (the primary path — a submitted `hint` command can't see the buffer it would help with, since Enter empties it), and a submitted **`hint` command** that expands on the *most recent error*. **No topic argument** (contextual only — `help <topic>` already owns explicit reference). Introduces a **tier-3 teaching layer**, deeper than the existing tier-1 (colour / error headline) and tier-2 (ambient one-liner; and the error `hint:`, which is shown **by default** since `Verbosity::Verbose` is the default — `messages short` is the opt-*out*); without it `hint` would just duplicate what's already on screen. Tier-3 content lives in the catalogue under `hint.cmd.<hint_id>` (per command form) and `hint.err.<class>` (per error/diagnostic class), each a structured `what`/`example`/`concept` block rendered via a new `note_hint*` family with `OutputStyleClass::Hint`. **Keyed per-form via a new `hint_ids: &[&str]` field on `CommandNode` mirroring `usage_ids`** (revised in Phase B): a per-*node* key proved too coarse — `add`/`drop`/`show`/`create` are each one node spanning many forms, and a live-input hint for `add 1:n relationship` must be specific to relationships; `hint_key_for_input_in_mode` reuses `usage_key_for_input_in_mode`'s form-word disambiguation, and covers the advanced-SQL forms whose `usage_ids` are empty. Not keyed off `help_id` (it is `None` on the advanced-SQL nodes purely to dedup the `help` list; that parallel gap is issue **#36**). **Clause-concept hints** (`on delete` actions, constraint slots, `with pk`, cardinality) are a recorded **deferred extension** (`hint.concept.<topic>`, issue **#37**) — per-form is the right tier-3 granularity, with position-awareness owned by tier-2 + the live `Next:` line. Two error routes share `hint.err.*`: pre-submit `diagnostic.*` read live from the walker (F1 path), runtime `translate_error` classes via stored `last_error_hint_key` (`hint` command / empty-F1). Adds `AppCommand::Hint`, a `HINT` grammar node + REGISTRY entry, the `hint_ids` field, and `last_error_hint_key`; F1 is a read-only overlay (buffer + completion memo untouched). **Content is the bulk of the work** (the mechanism is ~a day): **comprehensive for v1** — ~37 command forms + 9 runtime error classes + ~33 `diagnostic.*` classes ≈ 80 teaching blocks — authored **exemplars-first** (voice approved in this ADR's `/runda` review, then mass-authored in batches), enforced by a **comprehensiveness coverage test** (every node/error class has a key), with graceful fall-back to tier-2 if a key is ever missing. Forks user-chosen: two-surface model; **F1** (vs `?` / a chord); no-arg; comprehensive scope; exemplars-first. OOS: per-topic `hint <topic>` (rejected — overlaps `help`); always-on tier-3 (rejected — keeps ambient terse); non-`en-US` locales + success-command teaching (deferred); the `help`-side advanced-SQL gap (issue #36)